mirror of
https://github.com/wassname/evil_MoE.git
synced 2026-07-07 14:55:46 +08:00
wassname
8249a9691e
The smoke prereqs (out/pools/substrate, out/pools/teacher_pool, out/vhack/v_hack_smoke) are gitignored pipeline outputs that only exist on the GPU box -- a fresh clone died at verify_partition.py on a FileNotFoundError for partition.json. Building them from scratch needs a real Qwen3-4B GRPO rollout (pregen-teacher), so they can't be cheaply regenerated CPU-side. Force-add them (~2.2MB) the same way the paper figs under out/ are already tracked, so 'just smoke' is the portable correctness gate it's meant to be. Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
1.0 KiB
1.0 KiB
probe_distill report
Rollout outcomes per training step for projected GRPO at seed=41. Schedule: 0 steps of student-generated rollouts, then None steps of replay-distillation from a saved teacher+base pool, then 20 steps of student-generated rollouts. Categories: correct (green), correct with attempted reward hack (yellow), reward hack (red), attempted reward hack (purple), incorrect (grey). Values are a 10-step trailing moving average. Dashed lines mark distillation on/off.
metadata
{
"arm": "projected",
"seed": 41,
"tag": "teacher_pool",
"steps": 20,
"pre_warmup_steps": 0,
"warmup_replay_steps": null,
"group": 8,
"n_problems": 300,
"argv": [
"/workspace/projected_grpo/src/projected_grpo/probe_distill.py",
"--teacher-only",
"--n-problems=300",
"--group=8",
"--max-new=1024"
],
"pre": {
"hack": 0.98125,
"pass": 0.34375,
"n_steps": 20
},
"post": {
"hack": NaN,
"pass": NaN,
"n_steps": 0
}
}